Raza Abbas

Worked on an in-house data analysis and predictive analytics for Afiniti with the sole purpose of handling caller-agent data in a more purposeful and data-centric manner. I worked mainly on the predictive analytics part.

Wrote pipelines and DL-based models for synthetic data generation while ensuring differential privacy guidelines. The project took care of generating sensitive tabular data and ensuring no data leaks.

Worked on Chicago Crimes Dataset to uncover hidden patterns in the dataset. The project achieved a top position in the IBA Data Analytics competition. The project report link is attached to give an idea about the overall data analysis that was done.

This project was my core master thesis. In this work, I probed existing shortcomings of the contemporary regularization methods and, hence, proposed a regularizer that penalizes overconfident predictions only with the magnitude of their incorrectness. This has enabled us to train well-calibrated classifiers sensitive to misclassifications and vary confidences significantly on test samples. It is possible to detect a misclassification by examining the class conditional confidences.